After creating a perfectly generic account I started answering questions. However, I purposely chose to skip any questions that I thought would reveal my gender or my age—two demographics that might be key predictors of preferences. [I was going to make hunch work harder for my recommendations!] I didn’t stop @ 20 questions, probably answered about 60. The recommendations for me were only so-so and I noticed that all the recommendations rated highly for a majority of hunch users. OK. Predictable.
Of course what I really wanted to know is the magic (algorithms/rules/rules engine, etc) behind the connections. So…let’s try it out with more “use cases”! I made another generic account (not tied to a fb or twitter account) and answered THAY (Teach Hunch About You) questions as if I were my husband.
Results? I only answered 20 questions for my husband and was quite impressed with the returned suggestions. The cars (especially the muscle cars), and music groups (Rock) were dead on. The business books were pretty good too– 4 of the top 5 were excellent matches, in fact he brought one of them home for me to read several years ago. The worst category was dinner ideas – I don’t think there was one in the top 5 that he would be jazzed about. Comeon, herb-roasted vegetables? Really? But overall, why were his recommendations, so much more accurate than mine, and yet I answered 3 times as many questions for me? Maybe because when I answered for him, I readily answered questions that would reveal his gender or age?
So then I thought “ok, what if I answered questions like my DOG?” (a 12-year-old male pug) New generic log-in created, new questions. Results: Well first off, because it was so much fun, I ended up answering over 300 THAY questions for my dog! Some of my favorites that made me smile:
- Do you mind drinking out of public water fountains? No
- Do you like hot dogs? Yes (He’s old but he’s still interested!)
- Do you usually feel well-rested when you get up in the morning? Yes (even a picture of a dog with this question!)
- Word Association: I say bear, you say: Woof!
- Should doctor-assisted euthanasia be legal? No (think about it from a 12-yr-old dog’s point of view!)
How valid were the recommendations for my dog? Hard to say, really. For example under recommendations for pets - #1 was dogs, which is pretty predictable, but #2 was Rabbits, which was ranked above #3 Cats. [Is that predictable – for 12 year old males ? Perhaps] But this exercise in answering for others, led me to think of an educational use for hunch – that I doubt comes to mind readily: What if students were assigned to create generic accounts and answer questions based on what they knew of another person – a historical person, maybe Abraham Lincoln, or Madame Curie, or Amelia Earhart, or George Washington Carver? Or what if they did this in groups, for the same historical person, and then debated which group’s resulting recommendations were the best and why? Though I’m pretty sure that’s not what George had in mind as an educational use, it would be fun, and generate lots of collaboration, research, analysis and discussion, perhaps even an eventual discussion about stereotypes?
Oh, and another great educational use would be for students to create their own new topics and workshops, including the questions and answers. (If you don’t know what I’m referring to – look under “More” on the hunch menu.) Lots of applications and opportunities for discussion here, even about social media – because to get your topic in the “hunch corpus” it needs to have a certain number of thumbs up votes.
HOWEVER– getting to what I think is the best educational application of something “hunch-like”: It’s an example of something I’ve had rolling around in my head for a couple of years now. This is really just an adaptation of computer-adaptive testing. A computer adaptive test requires a large data-base of pre-calibrated items. The (very, very simplified here) theory is that a computer adaptive test can quickly zero-in on the subject’s ability (score) by iteratively asking harder questions until the subject answers incorrectly, and then easier questions until they answer correctly. The test “squeezes in” on the score/subject's "ability" much more quickly, and accurately with fewer questions than longer non-adaptive tests. There’s a lot of math and theory behind it too. Computer adaptive tests are generally created for dichotomously scored tests (where questions are scored either correct or incorrect). However there is also math and theory to calibrate and score polytomous questions – those that are more on a scale than a correct/incorrect basis, e.g, anxiety, self-efficacy, etc.
Now…what if we could do this for any web-mediated learning event? What if while experiencing such an learning event the user (or rather the data collected from the user’s interaction) was constantly being monitored, compared to previously “calibrated” data from all (and like) users, and the experience is automatically adapted to the user, based on this data – or, at least one or more recommendations is offered and the user selects how to continue? THIS in my opinion, is one way learning analytics will “earn it’s keep” so to speak. I know, more questions are raised than answered, but why not create a simplified proof of concept – oh, wait. That’s what hunch is :-) Ok, why not apply that to more “educational” material? Would that kill the “fun”? Would it be too controversial, raise too many privacy / or discrimination issues, take too long, be too much work, get no funding, etc. etc.? Caution usually doesn’t discover much of anything, especially anything very innovative.
Posts still to come: (eventually)
About this blog (including explanation of the name)
My Goals for the course
Introduction
That's an interesting idea. Maybe if we get two or three more ppl who are game we can setup some kind of prototype.
ReplyDeleteVicky Hennegan
Twitter: @eeus